1. What we collect
- Account information — email, display name and a hashed password, stored in our account system for authentication.
- Billing data — a customer ID from our payment processor. Card details are handled entirely by the processor; we never see or store your full card number.
- Usage metrics — translation minute counts and session metadata, used to enforce plan limits.
- Device data & Anti-Abuse — a one-way, non-reversible cryptographic hash of a hardware identifier (
device_hash) to prevent free-tier trial abuse; the raw identifier is never retained or uploaded. - Live audio streams — captured audio is transmitted transiently over encrypted WebSockets (
wss://) to Google Cloud (Gemini Live API) and Alibaba Cloud (Qwen Realtime API) for real-time translation. We do not store or retain your audio on our servers. Under enterprise API terms, transient audio is not used to train AI models. - On-device speaker-change detection — speaker-change labels (S1/S2) are computed 100% locally on your computer CPU using embedded
sherpa-onnxmodels, and the voice embeddings never leave your device. This marks when the voice changes; it does not identify who is speaking, cannot assign a name, and loses accuracy over overlapping speech or heavy music. - Local storage & Security — your browser stores a session token and language preferences; the desktop app stores settings, and any locally-held legacy BYOK API key is encrypted using Windows Data Protection API (DPAPI).
2. How we use it
To operate the real-time translation platform and your account, process billing through third-party payment systems, enforce usage quotas and prevent free-tier misuse, and send transactional email via Resend.
3. Third-party sub-processors
Data flows only to the sub-processors needed to deliver the service:
- Polar — payments and invoicing, as merchant of record. Card details are handled by Polar; we receive a customer ID and the order record.
- Our own server (PocketBase) — account, licence and usage records, on infrastructure we operate ourselves.
- Resend — transactional email: sign-up, purchase receipts and support replies.
- Google Cloud / AI Studio — Gemini, live audio translation.
- Alibaba Cloud / DashScope — Qwen, live audio translation.
We never sell or rent your data. The full list, with what each one receives, is on the sub-processors page.
4. Browser extension privacy
The Voxis browser extension uses tabCapture permissions to capture active tab audio in volatile memory for live translation overlays. Tab audio is processed in RAM only while translation is active and is never stored, profiled, or shared with advertisers.
5. Retention and deletion
What we keep, where it lives, and for how long:
| Data | Where it lives | How long |
|---|---|---|
| Account data (email, display name, password hash) | Our server | While your account is active |
| Usage events (minute counts, session metadata) | Our server | Kept with the account; deleted or anonymized when the account is deleted |
| Problem report (scrubbed log, diagnostics, and the transcript only if you tick the box) | Our server | While we investigate the case; deleted on request |
| Transcripts and recordings | Your own disk — they are never uploaded to our servers | Until you delete them |
| AI session summary | Sent to the model only when you press the button; the text comes back to your computer | Not stored on our servers |
| Live audio | Passes through the translation engine while the session runs | Not stored |
How to delete. Transcripts you delete yourself, from the saved-transcript list in the app; the file leaves your disk. For the account, write to [email protected]; we will delete or anonymize your account data within 30 days, except where legal or accounting obligations require retention.
6. Your rights
Depending on your jurisdiction (EU GDPR, UK GDPR, California CCPA, Türkiye KVKK), you have the right to access, correct, export, delete, restrict or object to the processing of your personal data. You may also lodge a complaint with your supervisory authority.
7. Security
Passwords are hashed with argon2/bcrypt; Legacy BYOK API keys are encrypted with hardware-bound Windows DPAPI; all transmission uses TLS/SSL (wss://, https://). On-device processing covers speech detection (VAD), speaker-change labelling and the free tier's local voice. On paid plans the captured audio streams continuously to the translation model while a session is live; on the free tier the stream is gated, so only detected speech is sent.
8. Children
The service is not directed at and may not be used by anyone under 18.
9. Contact & Data requests
Privacy questions and data protection requests: [email protected] or the contact page.